July 2025 UPDATE: We released version 1.1, adding almost 200k new queries 🎉🎉🎉.

v1.2 further adds the article titles as columns for convenience.

Use with:

country = "full" # "ar", "bo", ...
version = "1.2"
dataset = datasets.load_dataset("spanish-ir/messirve", country, revision=version)
print(dataset)

---

Dataset Card for MessIRve

MessIRve is a large-scale dataset for Spanish IR, designed to better capture the information needs of Spanish speakers across different countries.

Queries are obtained from Google's autocomplete API (www.google.com/complete), and relevant documents are Spanish Wikipedia paragraphs containing answers from Google Search "featured snippets". This data collection strategy is inspired by GooAQ.

The files presented here are the qrels. The style in which they are displayed makes them easier to explore, as it includes the full texts of documents and queries.

• For the conventional TREC-style topics and qrels files, refer to https://huggingface.co/datasets/spanish-ir/messirve-trec
• The corpus of documents that accompanies this dataset is https://huggingface.co/datasets/spanish-ir/eswiki_20240401_corpus

Dataset Details

Dataset Description

• Language(s) (NLP): Spanish
• License: CC BY-NC 4.0. The dataset should not be used for any commercial purpose.

Dataset Sources

• Repository: https://github.com/ftvalentini/MessirveSpanishIR
• Paper: MessIRve: A Large-Scale Spanish Information Retrieval Dataset

Uses

The dataset is meant to be used to train and evaluate Spanish IR models.

Dataset Structure

Data Instances

A typical instance of one subset of the dataset looks like:

{
  'id': 4918739,
  'query': 'a cual dedo se pone el anillo de compromiso',
  'docid': '956254#2',
  'docid_text': 'Pero desde hace cientos de años, se dice que la vena amoris pasa por el dedo anular izquierdo que conecta directamente al corazón (téngase en cuenta que la vena amoris no existe realmente). Tradicionalmente, es ofrecido por el hombre como regalo a su novia mientras o cuando ella accede a la proposición de matrimonio. Representa una aceptación formal del futuro compromiso.',
  'query_date': "2024-03-30",
  'answer_date': "2024-04-19",
  'match_score': 0.74,
  'expanded_search': false,
  'answer_type': 'feat_snip'
}

Data Fields

• id: query id
• query: query text
• docid: relevant document id in the corpus
• docid_text: relevant document text (excludes article title)
• query_date: date the query was extracted
• answer_date: date the answer was extracted
• match_score: the longest string in the SERP answer that is a substring of the matched document text, as a ratio of the length of the SERP answer
• expanded_search: if the SERP returned a message indicating that the search was "expanded" with additional results ("se incluyen resultados de...")
• answer_type: type of answer extracted (feat_snippet, featured snippets, are the most important)

Data Splits

We extract queries from Google's autocomplete API for 20 countries with Spanish as an official language, plus the United States. Equatorial Guinea was the only country left out because it doesn't have a Google domain.

Some API results were independent of the country-specific domain, many queries are not specific of any country. These are included under the country label none.

Queries from none were combined with the set of unique queries from all countries and included in the full subset. Unlike the country-specific sets, in the full set some queries can have multiple relevant documents because the same query may return different featured snippets in different country domains.

The dataset is partitioned into training and test queries in such a way that the Wikipedia article to which the paragraph belongs is present in only one of the splits. The partitioning was done by country, with about 20% of the articles assigned to the test set. Moreover, test instances always have match_score = 1 and expanded_search = False.

For further detail, such as statistics for each subset and split, see the paper.

Citation

BibTeX:

@inproceedings{valentini2025messirve,
    title = "{M}ess{IR}ve: A Large-Scale {S}panish Information Retrieval Dataset",
    author = "Valentini, Francisco  and
      Cotik, Viviana  and
      Furman, Dami{\'a}n  and
      Bercovich, Ivan  and
      Altszyler, Edgar  and
      P{\'e}rez, Juan Manuel",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1412/",
    doi = "10.18653/v1/2025.emnlp-main.1412",
    pages = "27740--27757",
    ISBN = "979-8-89176-332-6",
}

APA:

Francisco Valentini, Viviana Cotik, Damián Furman, Ivan Bercovich, Edgar Altszyler, and Juan Manuel Pérez. 2025. MessIRve: A Large-Scale Spanish Information Retrieval Dataset. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 27740–27757, Suzhou, China. Association for Computational Linguistics.